Explainable multi-step time series forecasting model

Abstract

This report presents an explainable multi-step time series forecasting model called the MSG-TCN. This model integrates multi-scale dilated convolutions (M), self-attention mechanisms (S) and gated activation units (G) on a Temporal Convolutional Network (TCN) network to effectively capture both local and global temporal patterns. We benchmark the MSG-TCN against four baselines, namely Autoregressive Integrated Moving Average (ARIMA), Long Short Term Memory (LSTM), a standard TCN, and the Galformer transformer model. For datasets, we used three major stock market indices (Standard & Poor’s 500 (S&P 500), Dow Jones Industrial Average (DJIA) and Nasdaq Composite Index (IXIC)). Experimental evaluations cover various forecast horizons (1-day, 10-day and 20-day), measuring performance via Mean Absolute Percentage Error (MAPE), Root Mean Squared Error (RMSE), Mean Absolute Error (MAE) and the coefficient of determination (𝑅2). Results demonstrate that MSG-TCN consistently attains comparable accuracy to state-of-the-art baselines while retaining a transparent structure amenable to interpretability. We also integrated local interpretable Shapley additive explanation methods into MSG-TCN, allowing us to visualize the key features that influence the model’s decisions and improve its interpretability. By merging precise multi-step forecasting with enhanced interpretability, our project delivers a robust, scalable solution for high-stakes decision-making while driving continued improvements in the financial forecasting domain.